devcontainer-mcp

Give your AI agent its own dev environment — not yours.

devcontainer-mcp is an MCP server that lets AI coding agents create, manage, and work inside dev containers across three backends: local Docker, DevPod, and GitHub Codespaces. The agent builds, tests, and ships code in an isolated container — your laptop stays clean.

Works with GitHub Copilot, Claude, Cursor, opencode, and any MCP-compatible client.

The Problem

When AI agents write code, they need to run it somewhere. Today that means your host machine:

• 🔴 Host contamination — agents install packages, modify PATH, leave behind build artifacts
• 🔴 "Works on my machine" — agents assume your local toolchain matches production
• 🔴 No isolation — one project's dependencies break another
• 🔴 Security risk — agents run arbitrary commands with your user privileges
• 🔴 Hardware constraints — you're limited to your local machine's resources

The Solution

The devcontainer spec already defines reproducible, container-based dev environments. Every major project ships a .devcontainer/devcontainer.json. But AI agents can't use them — until now.

devcontainer-mcp exposes 45 MCP tools that let any AI agent:

1. Spin up a dev container from any repo — locally, on a cloud VM, or in Codespaces
2. Run commands inside the container — builds, tests, linting, anything
3. Manage the lifecycle — stop, restart, delete when done
4. Authenticate against cloud providers — GitHub, AWS, Azure, GCP — without ever seeing a raw token

Agent: "Let me build this project..."
  → auth_status("github") → picks account
  → codespaces_create(auth: "github-you", repo: "your/repo")
  → codespaces_ssh(auth: "github-you", codespace: "...", command: "cargo build")
  → ✅ Built in the cloud. Your laptop did nothing.

Quick Install

Linux / macOS

curl -fsSL https://raw.githubusercontent.com/aniongithub/devcontainer-mcp/main/install.sh | bash

Windows (via WSL)

Invoke-RestMethod https://github.com/aniongithub/devcontainer-mcp/releases/latest/download/install.ps1 | Invoke-Expression

How it works: The binary runs inside WSL; MCP clients on Windows launch it via wsl ~/.local/bin/devcontainer-mcp serve. The stdio transport works transparently across the WSL boundary. WSL 2 is required — install it with wsl --install if you haven't already.

Backend CLIs (devpod, devcontainer, gh) are detected at runtime — if one is missing, the MCP server returns a helpful error with install instructions.

Binaries available for linux-x64, linux-arm64, darwin-x64, and darwin-arm64.

Host Protection & Opt-Out

The installer configures two agent hooks (Claude Code & GitHub Copilot CLI) that activate only in a directory containing .devcontainer/devcontainer.json:

• devcontainer-guard (PreToolUse) — blocks shell/command execution on the host so the agent routes builds, tests, and runs through the MCP tools instead. Host-safe commands (git, gh) are allowlisted, and both hooks fail open — if jq is missing or a payload can't be parsed, commands are allowed rather than blocked.
• devcontainer-skill-loader (SessionStart) — injects the SKILL.md usage guide as context.

Turning it off for a repo

Sometimes a project ships a stale or unmaintained .devcontainer that you don't actually want to route work through. To disable both hooks for that repo, drop a marker file at the repo root:

touch .devcontainer-mcp-disable

When .devcontainer-mcp-disable is present, the guard allows host commands through and the skill-loader skips context injection — the agent works locally as if no devcontainer were declared. Commit it (or add it to .gitignore for a local-only opt-out).

One-off bypass: to run a single host command without disabling the guard, include USER_CONFIRMED_HOST_OPERATION=1 anywhere in the command.

Architecture

graph TD
    A[AI Agent / MCP Client] -->|stdio JSON-RPC| B[devcontainer-mcp]
    
    subgraph "devcontainer-mcp"
        B --> C[33 MCP Tools]
        C --> D[Auth Broker]
        C --> E[devcontainer-mcp-core]
    end
    
    D -->|opaque handles| C
    E -->|subprocess| F[DevPod CLI]
    E -->|subprocess| G[devcontainer CLI]
    E -->|subprocess| H[gh CLI]
    E -->|bollard API| I[Docker Engine]
    
    F --> J[Docker / K8s / Cloud VMs]
    G --> K[Local Docker]
    H --> L[GitHub Codespaces]

Three Backends, One Interface

Auth Broker

The agent never sees raw tokens. Instead:

1. auth_status(provider) — list available accounts and scopes
2. auth_login(provider, scopes?) — initiate login, opens browser, handles device codes
3. auth_select(id) — switch the active account
4. auth_logout(id) — revoke credentials

Codespaces tools require an auth handle (e.g. "github-aniongithub"). The MCP server resolves it to the real token on each call via the CLI's native keyring.

Supported providers: GitHub, AWS, Azure, GCP, Kubernetes

MCP Tools (46 total)

Auth (4 tools)

DevPod (19 tools)

devcontainer CLI (12 tools)

GitHub Codespaces (11 tools) — require auth handle